Niket Patel

Most people with “AI/ML” in their bios don’t even know a real symmetric matrix always has real eigenvalues.

Can we train LLMs with RL using the same next token prediction loss as pre-training? (yes) We conduct a study on (log)prob rewards and show they give a simple way to bridge verifiable and non-verifiable settings with a single reward, broadly applicable for fine-tuning LLMs.

Can a model learn to break its own reasoning plateau? In our new paper, we show that LLMs can be taught with meta-RL to generate their own "stepping stones" that kickstart learning on hard math problems (0/128 success rate) where direct RL fails. Paper 📝: arxiv.org/abs/2601.18778 Blog post 🌐: ssundaram21.github.io/soar/ (1/n)

Want to teach AI agents to use apps like humans? Get started with digital agents research using OpenApps, our new Python-based environment.

I will be recruiting 1-2 PhD students at @NYUDataScience or @NYUCourant CS to work on Machine Learning & applications in NYU's vibrant top ML ecosystem. Check Google Scholar to see our latest research interests. Interested? Please mention my name in your application. Deadl. 12/12

New paper alert! When LLMs write long responses to open-ended questions, how can you tell if they're hallucinating - without checking every single claim? We found something interesting hidden *inside embeddings*. arxiv.org/pdf/2510.21891 With @timrudner @KempeLab @NYUDataScience Thread 🧵 1/5

Current GRPO wastes compute on negative groups — when all K samples are wrong, you get zero gradient despite full generation cost. We propose a principled fix by bridging reward modeling and policy optimization: 👉 Penalize highly confident wrong answers more to create signal.🧵

RL has led to amazing advances in reasoning domains with LLMs. But why has it been so successful, and why does the length of the response increases during RL? In new work, we introduce a framework to provide conceptual and theoretical answers to these questions.

Measuring information content in bits is very useful. Information theory made digital communication, cryptography and machine learning possible. But information is not just a quantity: it also has a shape. (1/6)

@ziv_ravid A podcast where you only interview graduate students called "grad student descent"

We want to start a podcast about cutting-edge AI research and technical breakthroughs. Need a catchy name! What would you call it? The one who suggest the best name will be our guest 🥳

anyway, none of this is possible because the data processing inequality

Our new ICML 2025 oral paper proposes a new unified theory of both Double Descent and Grokking, revealing that both of these deep learning phenomena can be understood as being caused by prime numbers in the network parameters 🤯🤯 🧵[1/8]

Hi, happy to share that our paper "Layer by Layer: Uncovering Hidden Representations in Language Models" was accepted to ICML as a spotlight! 🥳🥳🥳 Oscar Skean, @rarefin15, DanZhao, Jalal Naghiyev and @ylecun

@miniapeur Berkeley Problem's in Mathematics is an option

I'm looking for mathematics book full of problems to solve and, ideally, their solutions. I want to start solving at least one problem a day and it's easier if I can just browse through a book. It should be at least undergrad level. I am open to a book covering several topics or specific topics such as linear algebra, probability, analysis, statistics, etc.

@kuanhenglin @UCLAComSci @sicheng_mo @zhoubolei Congrats!!

Honored to be selected as a CRA Outstanding Undergraduate Researcher Award Finalist and to be featured on the @UCLAComSci website! Could not have done this without the mentorship of @sicheng_mo and @zhoubolei along the way :D cs.ucla.edu/ucla-cs-underg…

@sfrei_ @GoogleDeepMind Looking forward to seeing what you work on!

Job update: I've joined @GoogleDeepMind as a research scientist! I'll be working from the SF office. Super excited!

Entropy will be a key word in 2025 👀

@ziv_ravid not sure I think the booth is closed, but I saw someone with a brick at some point

@NiketPatel91154 where?!?!

People at NeurIPS, does anyone have extra swag from the conference? I went to the booths, but apparently they were already gone. My kids will kill me 😱

CDS Research Scientist Ravid Shwartz-Ziv @ziv_ravid is experimenting with open scientific collaboration through a Discord server, resulting in projects on neural network compression & language model layers that have already produced a @NeurIPSConf paper. nyudatascience.medium.com/open-research-…